JPH11237894A - Method and device for comprehending language - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice interactive system with which a voice having the non-clear partition of speaking can be comprehended in real time and immediate responding is enabled by speaking at free timing. SOLUTION: When any word is inputted, while referring to a dictionary 250, the semantic description of that word is found by an interactive context managing part 210, a new interactive context candidate is provided together with respective interactive context candidates provided up to that time point, a rule application part 220 obtains a further new interactive context candidate by applying language comprehension rules 110 to these interactive context candidates, a priority calculating part 230 calculates priority by adding these new interactive context candidates to the source interactive context candidates, and the interactive context managing part 210 outputs the interactive context candidate of the highest priority. This processing is repeated each time a word is inputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a language comprehension method and a language comprehension device for comprehending the language of an input utterance and outputting a dialogue context candidate. The present invention relates to a language understanding method and a language understanding device used in a voice dialogue system or the like which inputs a character string as a character string to a computer through a computer and processes the meaning of the character string by computer.

[0002]

2. Description of the Related Art Conventional language understanding techniques have been applied to those in which a break in a sentence or utterance is clear. When analyzing written words, use punctuation marks as sentence breaks, analyze the text between punctuation marks using language understanding rules, calculate their semantic contents, and the context calculated from the preceding sentence. A method of calculating new contextual information along with the information was used. In the case of analysis of spoken language, the user can press a button or the like at the time of uttering to check the utterance section, or use a pause to separate the utterance, perform a linguistic analysis of the immediately preceding utterance section after ending the utterance, and determine the semantic content. A method of calculating and calculating context information was used.

[0003]

In spoken language, the unit of sentence is not clear. Sometimes a single word indicates intent, and sometimes a long utterance is spoken. If a break such as a pause is used, the input unit is not determined until the pause appears, so that the meaning of the utterance cannot be grasped and a high-speed reaction cannot be performed. Further, there are cases where the user speaks a lot of matters without interruption, and in the middle of that, there is also a case where the user has to react appropriately such as comprehending the meaning and responding to each other, such as asking and asking a question. Even if you try to find utterance breaks using language comprehension rules from continuous utterances, the rules are ambiguous and can be understood in various ways, so it is unclear which one to use There is a problem.

[0004] It is an object of the present invention to provide a language understanding method and apparatus capable of realizing speech understanding in real time even if the utterance break is not clearly understood.

[0005]

According to the language solving method of the present invention, every time a word is input, a meaning description of the word is obtained by referring to a dictionary, and the dialogue obtained up to that time is obtained. A new dialogue context candidate is determined along with each of the context candidates,
A language understanding rule is applied to these dialogue context candidates to obtain new dialogue context candidates, and the obtained dialogue context candidates are combined with the original dialogue context candidates to calculate a priority. The feature is to output a dialogue context candidate having a high probability.

[0006] Further, the number word interpreting apparatus of the present invention comprises means for holding dialogue context candidates obtained up to a previous time point, and when a word is input, a semantic description of the word by referring to a dictionary. Means for determining a new dialog context candidate along with each of the dialog context candidates obtained up to that point;
A means for applying a language understanding rule to the obtained dialogue context candidate to obtain a new dialogue context candidate, and a means for calculating a priority by combining the obtained dialogue context candidate with the original dialogue context candidate. It is characterized in that the processing of these means is repeated every time a word is input, and a dialogue context candidate with the highest priority is output at each time.

[0007]

FIG. 1 shows a processing flow diagram of an embodiment of a language understanding method according to the present invention, which is executed with the aid of a so-called computer. When a word is input (step 110), first, referring to a dictionary (word dictionary), a description of the meaning of the input word (part of speech, semantic classification, etc.)
(Step 120). Next, this semantic description is
A new combination of dialogue context candidates is obtained by combining with each of the dialogue context candidates obtained up to the time of the preceding word input (step 130). Thereafter, language understanding rules are applied to these dialogue context candidates, and a new combination of dialogue context candidates is obtained from the result (step 140). Then, the priority is calculated by combining these new dialogue context candidates with the original dialogue context candidates (step 150), and the dialogue context candidate having the highest priority is output (step 160). Hereinafter, every time a word is input, the processing of steps 110 to 160 is repeatedly executed.

As described above, in the language understanding method of the present invention, when there are a plurality of ways of applying the language understanding rules to the input utterance, a plurality of candidates for the dialogue context obtained as a result of applying each method are determined. After storing these and calculating their priorities, the highest priority one is checked and output.
As a result, even if the utterance break is not clearly understood, it is possible to obtain the context with the highest priority at each point in time, so that it is possible to output in real time.

Next, FIG. 2 shows a configuration diagram of an embodiment of the language understanding device according to the present invention. The language understanding device 200 includes a dialogue context management unit 210, a rule application unit 220, a priority calculation unit 230, a dialogue context candidate holding unit 240, a dictionary (word dictionary) 250, and a language understanding rule 260. The conversation context candidate holding unit 240 holds the conversation context candidates obtained up to the previous time. When a word is input, the dialogue context management unit 210 obtains a semantic description of the input word by referring to the dictionary 250, and obtains the meaning description of the input word until the time of the previous word input held in the dialogue context candidate holding unit 240. By combining with each of the dialogue context candidates obtained in (1), a new combination of dialogue context candidates is obtained and sent to the rule application unit 220. The rule application unit 220 attempts to apply the language understanding rule 260 to each of the dialogue context candidates sent from the dialogue context management unit 210, and if it succeeds, obtains new dialogue context candidates as a result. The obtained dialogue context candidate is sent to the priority calculation unit 230 together with the original dialogue context candidate. The priority calculation unit 230 calculates the priority of each of the dialog context candidates sent from the rule application unit 220 and returns the priority to the dialog context management unit 210. The conversation context management unit 210 writes each of the conversation context candidates with the priority returned from the priority calculation unit 230 into the conversation context candidate holding unit 240, and stores the conversation context candidate with the highest priority at that time. Output.

FIG. 3 shows an example of the present language understanding device 200.
1 is a configuration diagram of a voice interaction system to which the present invention is applied. In FIG. 3, when a speech to be processed is input to the speech recognition apparatus 300, the speech is sent to the language understanding apparatus 200 every time a word in the speech is recognized. In the language understanding device 200, as shown in FIG. 2, every time a word is input, a dialogue context is obtained using the dictionary 250 and the language understanding rule 260, and is sequentially output.
This is sent to the response planning device 400. The response planning device 400 determines whether or not a response is required according to the context information, plans an appropriate response if necessary, and generates and outputs a response message using the speech synthesis device 500.

Here, as a specific example, an understanding of an input of "first conference room, 2:30" (hereinafter, utterance 1) during a dialogue for reserving a conference room will be considered. The voice recognition device 300 outputs a message “1st meeting room, 2:30, 30 minutes” from the voice input.
Each time one word is recognized, the language understanding device 200
Send to It is assumed that the dictionary 250 and the language understanding rule 260 in the language understanding device 200 are as shown in FIGS. 4 and 5, respectively. Each line in FIG. 5 corresponds to a language understanding rule. For example, utterance (meeting room setting) ← noun (meeting room) indicates that a noun whose semantic classification is “meeting room” can be interpreted as an utterance instructing “meeting room setting”. A noun (time) ← noun (time) noun (minute) is a sequence of nouns whose semantic classification is “time” and a series of nouns whose semantic classification is “minute”. "Indicates that it can be interpreted as a noun.

First, when the word "first conference room" is input, the dialogue context management unit 210 forms the following dialogue context candidate (1). Note that the conversation context candidate holding unit 240 is empty at this time. (1) Priority: 0 Understanding record: Noun (meeting room, first meeting room) A dialogue context candidate is composed of a priority and an understanding record, and the initial value of the priority is 0. The comprehension record records the state of comprehension of the conversation so far. Here, a semantic description “noun (meeting room, first meeting room)” is obtained from the dictionary item of “first meeting room” in FIG. 4 and recorded.

Next, the dialogue context candidate (1) is sent to the rule application unit 220. The rule application unit 220 applies the language understanding rule of FIG. 5 to the understanding record of the dialogue context candidate (1).
In this case, the rule (A) in FIG. 5 can be applied, and the following new dialog context candidates (2) and (3) can be obtained.

(2) Priority: 0 Understanding record: Noun (meeting room, first meeting room) (3) Priority: 0 Understanding record: utterance (meeting room setting, first meeting room) Dialogue context candidate (2) Is a rule in which no rule is applied to the dialogue context candidate (1), and a dialogue context candidate (3) successfully applies the rule (A) in FIG.

These two dialogue context candidates (2) and (3)
Is sent to the priority calculation unit 230. Here, the priority of the dialogue context candidate (3) to which the rule is applied is increased, and the following (3) 'is obtained. (3) 'Priority: 10 Understanding record: utterance (meeting room setting, first meeting room) The priority depends on the rule used and the state of the understanding record at that time. In this example, each time one rule is applied,
Assume that it is increased by "10". This dialogue context candidate (3) '
Is sent from the dialogue context management unit 210 to the response planning device 400.

Next, the dialogue context candidates of the dialogue context management unit 210 when "2 o'clock" is input are as follows. (4) Priority: 0 Understanding record: Noun (meeting room, first meeting room), noun (time,
(2) (5) Priority: 10 Understanding record: utterance (meeting room setting, first meeting room), noun (time, 2:00).

The dialog context candidates (4) and (5) are as follows through the rule application unit 220 and the priority calculation unit 230. (6) Priority: 10 Understanding record: Noun (meeting room, first meeting room), utterance (time setting, 2 o'clock) (7) Priority: 10 Understanding record: uttering (meeting room setting, first meeting room) , Noun (time, 2 o'clock) (8) Priority: 20 Understanding record: utterance (meeting room setting, first meeting room), utterance (time setting, 2:00) The dialogue context candidate (6) is a dialogue context candidate ( The rule (B) of FIG. 5 is applied to 4). Dialogue context candidate (7)
Is a rule in which no rule is applied to the dialogue context candidate (5), and the dialogue context candidate (8) is a dialogue context candidate (5).
In addition, the rule (B) of FIG. 5 is applied. Those to which no rule is applied to the dialogue context candidate (4) are discarded because they have low priority. As a result, the dialogue context candidate (8) is sent from the dialogue context management unit 210 to the response planning device 400.
Sent to

Next, when "30 minutes" is input, the dialogue context management unit 210 makes the following dialogue context candidates (9),
(10) and (11) are formed. (9) Priority: 10 Understanding record: utterance (meeting room setting, first meeting room), noun (time, 2 o'clock), noun (minute, 30 minutes) (10) Priority: 10 understanding record: noun (meeting Room, first meeting room), utterance (time setting, 2:00), noun (minute, 30 minutes) (11) Priority: 20 Understanding record: utterance (meeting room setting, first meeting room), utterance (time setting) , 2 o'clock), nouns (min, 30 min).

These dialogue context candidates (9), (10),
(11) is as follows through the rule application unit 220 and the priority calculation unit 230. (12) Priority: 20 Understanding record: utterance (meeting room setting, first meeting room), utterance (time setting, 2:00), noun (minute, 30 minutes) (13) Priority: 20 Understanding record: utterance ( Meeting room setting, first meeting room), noun (time, 2:30) (14) Priority: 30 Understanding record: utterance (meeting room setting, first meeting room), utterance (time setting, 2:30) Here, the dialogue context candidate (12) is obtained by applying no rules to the dialogue context candidate (11). The dialogue context candidate (13) is obtained by applying the rule (C) of FIG. 5 to the dialogue context candidate (9), and the dialogue context candidate (14) is obtained by applying the rule (B) of FIG. It is. As a result, the dialogue context candidate (14) is sent from the dialogue context management unit 21 to the response planning device.

That is, after "2 o'clock" is input, the interpretation that the time setting is "2 o'clock" has the highest priority, but "3 o'clock" has the highest priority.
After "0 minutes" is input, the time setting is "2:30"
Is the highest priority. In this way, the interpretation with the higher priority may change sequentially. In this way, the context with the highest priority at each time point is obtained.

[0021]

As described above, by using the language understanding method and apparatus of the present invention, it is possible to understand in real time a speech whose utterance is not clearly separated. As a result, even if the user speaks at a free timing, if a response is required, it is possible to create a voice interaction system or the like that can immediately respond.

[Brief description of the drawings]

FIG. 1 is a processing flowchart of an embodiment of a language understanding method according to the present invention.

FIG. 2 is a configuration diagram of one embodiment of a language understanding device of the present invention.

FIG. 3 is a system configuration diagram of an application example of the language understanding device of the present invention.

FIG. 4 is a diagram illustrating an example of a dictionary.

FIG. 5 is a diagram illustrating an example of a language understanding rule.

[Explanation of symbols]

 210 Dialogue Context Manager 220 Rule Applyer 230 Priority Calculator 240 Dialogue Context Candidate Holder 250 Dictionary 260 Language Understanding Rules

Claims (2)

[Claims] 1. A language understanding method for performing language understanding of an utterance and outputting a dialogue context candidate. Each time a word is input, a meaning description of the word is obtained by referring to a dictionary. A new dialogue context candidate is obtained along with each of the dialogue context candidates obtained up to the point in time, language understanding rules are applied to these dialogue context candidates, and further new dialogue context candidates are obtained. A language understanding method comprising: calculating a priority of a dialogue context candidate that has been combined with an original dialogue context candidate; and outputting a dialogue context candidate having the highest priority. 2. A language understanding device for performing language understanding of an utterance and outputting dialogue context candidates, comprising: means for holding dialogue context candidates obtained up to a previous time; Means for obtaining a semantic description of the word by referring to the dictionary, and obtaining new dialogue context candidates together with each of the dialogue context candidates obtained up to that time; Means for obtaining a new dialogue context candidate by applying an understanding rule; and means for calculating a priority by combining the obtained dialogue context candidate with the original dialogue context candidate. A language understanding apparatus characterized by repeating each time a word is input, and outputting a conversation context candidate having the highest priority at each time.